🔍
A Deep Learning Based Anti-aliasing Self Super-resolution Algorithm for Magnetic Resonance Images
Case ID:
C15254
Report of Invention:
4/9/2018
Web Published:
7/12/2019
Unmet Need
MRI has been desired and used for many clinical applications. When obtaining high resolution images, there is a requirement for an adequate signal-to-noise ratio that necessitates a long acquiring time. This makes image acquisition costly and susceptible to many motion artifacts, or noised caused that affects image quality. A common way to solve this problem has been applied to acquire MR images that have both good in-plane resolution and a large slice thickness. The problem with this method is aliasing is introduced in the through-plane direction, causing high frequency artifacts that can’t be removed through conventional methods. There is a need for the successful use of super-resolution, which has the idea to remove aliasing artifacts and improve spatial resolution. Super resolution methods are mostly learning-based and require external training data, which unfortunately due to scanner limitations, is often not available. Therefore, there is a need to overcome this issue to allow for super resolution algorithms to actually be implemented. There is a need for anti-aliasing and super resolution algorithms will allow for high resolution MR images to be taken in a cost effective and efficient manner, and it must be better than current self-super-resolution algorithms.
Technology Overview
The inventors have proposed an anti-aliasing and self-super resolution algorithm that is self-learning and is able to produce these high resolution MR images without the need for external data. This solves the issue of training the algorithm to acquire these images, as that was one of the main problems hindering this research. The algorithm consists of three steps: 1.) A Self-Anti-Aliasing (SAA) and Self-Super-Resolving (SSR) deep network is created. 2.) This network is created to be able to be applied in several different orientations, allowing for it to take advantage of the fact that high resolution information can be acquired in-plane slices. 3.) Finally, recombining the multiple orientations found in step 2. This SAA+SSR algorithm can be applied to a diverse collection of MR data, all neither modified nor preprocessed. This solution can be applied to many different areas, such as research, diagnosis, and prognosis. By being able to effectively obtain information about the high frequencies visually, researchers and clinicians alike can begin to tackle some of the world’s health problems from a completely new perspective.
Stage of Development
The inventors have developed a prototype of the algorithm are currently testing and checking the improvement of the algorithm compared to other competing methods.
Patent Information:
Title
App Type
Country
Serial No.
Patent No.
File Date
Issued Date
Expire Date
Patent Status
MACHINE LEARNING PROCESSING OF CONTIGUOUS SLICE IMAGE DATA
PCT: Patent Cooperation Treaty
United States
17/274,901
11,741,580
3/10/2021
8/29/2023
8/16/2040
Granted
MACHINE LEARNING PROCESSING OF CONTIGUOUS SLICE IMAGE DATA
CON: Continuation
United States
18/350,816
12,067,698
7/12/2023
8/20/2024
9/13/2039
Granted
Direct Link:
https://jhu.technologypublisher.com/technology/34950
Inventors:
Category(s):
Technology Classifications > Computers, Electronics & Software > Image Processing & Analysis, Technology Classifications > Computers, Electronics & Software > Machine Learning, Technology Classifications > Medical Devices > Imaging,
Get custom alerts for techs in these categories/from these inventors:
Subscribe for JHTV Updates
For Information, Contact:
Andrew Wichmann
wichmann@jhu.edu
410-614-0300
Save This Technology:
Bookmark this page
Download as PDF
JHTV Home
|
Search
|
Login/Subscribe
2017 - 2022 © Johns Hopkins Technology Ventures. All Rights Reserved. Powered by
Inteum